A collection of reusable skills for AI agents that extend capabilities with specialized knowledge, workflows, and tools.
Skills are reusable capabilities that help AI agents accomplish specific tasks more effectively. This repository contains skills that can be installed and used with the skills.sh ecosystem.
Install the skills CLI:
npm install -g @vercel/skills
# or
npx skillsSkills can be installed from git repositories, URLs, or local paths:
# Install from GitHub repository
npx skills add https://github.com/username/repo
# Install from GitHub shorthand (user/repo)
npx skills add username/repo
# Install from local directory
npx skills add ./path/to/skill-directory
# Install from this repository
npx skills add RussBrown00/agent-skills# List installed skills
npx skills list
# Check for updates
npx skills check
# Update all skills
npx skills update
# Initialize a new skill template
npx skills init my-new-skillOnce installed, skills provide procedural knowledge that enhances what your AI agent can do. Each skill contains:
- SKILL.md: Instructions and guidance for using the skill
- Optional resources: Scripts, references, and assets for complex tasks
Skills are automatically loaded by compatible AI agents when the skill's trigger conditions are met.
Interrogate a plan before you execute it, roast the code after you write it, and drive review findings to zero.
New to these? WORKFLOW-SKILLS.md is a one-page tour of
sniff-my-plan,roast-my-plan,roast-my-code,dead-horse, andeat-the-frog— the what and why of each, with links to the full guides. Share that single link to introduce the set.
The lightest plan gate — a pre-execution intent-alignment sniff test. Reconstructs the plan's intent as a five-part statement (Goals, Why, How-it-achieves-the-goals, Non-Goals, and Success/Done), filling what the plan supports and asking only where a section is too thin — questions batched by section, each with a proposed default, endable with "good enough." Then it states its understanding back and starts an alignment conversation, flagging where the reconstructed intent and the plan's actual content diverge (each cited to a plan line) so you can adjust before committing to the work. Checks what and why — not executability (roast-my-plan) or the agent's silent confidence (dead-horse) — and is advisory and ephemeral: screen-only, it writes no file and never edits the plan. Verdict is ALIGNED / MISALIGNED. If no file is given, it resolves the plan from context (a plan touched this session, an inline plan in the conversation, or .omc/plans/{TICKET}/ from the branch).
Triggers: /sniff-my-plan {file?}, "sniff my plan", "sniff-test this plan", "do we agree on the goal", "gut-check this plan", "is this plan rich enough"
Read more: sniff-my-plan/README.md — the five-part statement, the "rich enough" bar, and the alignment conversation.
Pre-execution executability gate for an existing plan file, delivered like a hostile principal engineer. Answers one question with evidence: could a competent engineer execute this plan end-to-end without asking anything? It hunts contradictions, blank spaces, ambiguity, bad architecture, and unnecessary complexity across six review lenses, verifies every anchor the plan names against the codebase, and converges on a falsifiable four-gate rubric — every reference resolves, zero open blocking questions, acceptance criteria are executable, no internal contradictions — for a READY / NOT READY verdict. Anything the plan, its referenced docs, or the codebase can settle is verified directly rather than asked; the question loop is uncapped (delivered in small groups) with "accept residual risk" as the brake. Runs in one of two modes chosen at entry: Report (writes a {plan}-roasted.md sibling, never touches the plan) or Fix-live (edits the plan in place as you approve each finding). Read-only against the codebase in both modes.
Triggers: /roast-my-plan {plan-file} {instructions?}, "roast this plan", "roast my plan", "tear this plan apart"
Read more: roast-my-plan/README.md — the four-gate rubric, the two modes, and how the roast works.
The sister of roast-my-plan, on the code side: a delta-focused, forge- and LLM-agnostic hostile code review. Point it at a target — omit for the whole current branch (incl. uncommitted) vs the trunk merge-base (confirmed with you first), or pass in-progress, HEAD, a PR number (fetched via gh/glab, reviewed locally), a commit SHA, or a branch — and it turns an embedded Linus-style pedantic reviewer loose across three context-blind lenses (correctness & concurrency, tests & verification, API-contract/security/design), fanned out as parallel subagents or a serial multi-pass when the client has none. Every finding cites a changed line (grep-verified against the diff) and is proven by running code where feasible; graded High / Medium / Low with a soft one-line verdict. For a PR target it suppresses findings other reviewers already raised; for local targets it presents everything. Two modes chosen at entry: Report (writes findings, touches nothing) or Fix-live (edits the source in place as you approve each finding, rendering the report against original line numbers first). Read-only against version control — never commits or pushes.
Triggers: /roast-my-code [target] {instructions?}, "roast this code", "roast my code", "roast this diff", "roast this branch", "roast this PR", "tear this code apart"
Read more: roast-my-code/README.md — the target forms, the three lenses, PR-only suppression, and the two modes.
Probes the quality of the agent's own confidence in a plan. Instead of asking what it is unsure about, the agent enumerates what it is certain of — the assumptions it filled in without noticing — states its answer to each, and asks you to confirm or correct them. Every correction is researched, turned into a specific proposed plan edit, and applied on approval. The miss rate is scored and clustered to show where its certainty is worthless, and rounds repeat for as long as you want them.
Triggers: /dead-horse {plan-file?} {count:15} {instructions?}, "check your assumptions", "what are you assuming", "probe your confidence", "beat this to death"
Read more: dead-horse/README.md — the problem it solves, worked examples, and how it differs from roast-my-plan.
Works a list of review findings one issue at a time, hardest first by severity x blast radius. Keeps an untracked, restart-safe tracker file, snapshots the git baseline, restates each issue in full, proposes a fix before editing anything, and closes nothing until both you and the agent agree it is resolved. Handles code, plan-file, and doc findings.
Triggers: /eat-the-frog {optional-instructions}, "step through these", "work through these one at a time", "hardest first", "resume eat-the-frog" — and it offers itself, once and unprompted, whenever a review or check has just left two or more findings unresolved
Read more: eat-the-frog/README.md — the problem it solves, the six-phase loop, and worked examples.
Interacts with live Hunk diff review sessions via CLI. Inspects review focus, navigates files and hunks, reloads session contents, and adds inline review comments.
Triggers: a running Hunk session, or wanting to review diffs interactively
Commit hygiene and history reporting.
Creates standardized commits, amends, and squash-merges from the current repo state, using conventional-commit-style summaries with multi-line bullet bodies.
Triggers: "commit", "stage and commit", "amend", "generate a commit message", "squash merge to <hash>" — including when a git subagent asks
Pushes the current branch to origin and opens a draft PR/MR, with the title derived from the commit chain and the body listing every commit since the configured base branch. Forge (GitHub/GitLab) is auto-detected from the origin URL. Pass squash to enable auto-squash-merge on GitHub, or --dry-run to preview without changes.
Triggers: /git-push-pr, "push and open a PR", "create a pull request for the current branch"
Deployment summary of commits from a given hash to HEAD, as an indented bullet list with the project folder as the header and one line per commit.
Triggers: /git-summary <hash>, "deployment notes", "release notes", "commits since a hash"
Language-level conventions for writing and testing JS/TS.
Project-specific JavaScript and TypeScript conventions for code design, module layout, control flow, data flow, async, and error handling.
Triggers: writing or editing .js, .jsx, .ts, .tsx, .mjs, or .cjs files; defining functions; organizing modules; refactoring
Zod schema practice — extending base schemas for specific contexts, deriving types, and handling validation boundaries.
Triggers: writing or reviewing TypeScript that uses Zod, or type assertions appearing near Zod validation
Best practices for writing JavaScript and TypeScript tests using Jest, including mocking strategies, test structure, and common patterns.
Triggers: writing or modifying Jest tests, mocking modules, structuring test suites
React-specific conventions, upgrades, and routing. Layers on top of the JavaScript & TypeScript skills.
Project-specific React conventions for component shape, hooks, props, state, effects, refs, and rendering.
Triggers: writing or editing .jsx or .tsx files, defining components, writing or modifying hooks, working with JSX
Upgrades React 17 applications to React 18 safely and incrementally — dependency updates, ReactDOM.render to createRoot/hydrateRoot, automatic batching, StrictMode fallout, test updates, and React 19 preparation.
Triggers: working in a React 17 codebase, resolving React 18 deprecation warnings, finishing a partial upgrade, validating React 18/19 compatibility
Stub — not yet written. This directory still contains the unmodified skills init scaffold: the frontmatter description and the body are both placeholder text.
Triggers: none. It will not meaningfully activate until it is filled in.
Producing prose, and producing skills.
Prose that reads as human rather than LLM-generated — comments, commits, posts, copy, and PR descriptions.
Triggers: drafting responses, social copy, technical docs, inline comments, commit messages, blog posts
Guide for creating an effective new skill, or updating an existing one, that extends an agent's capabilities with specialized knowledge, workflows, or tool integrations.
Triggers: "create a new skill", "update this skill", authoring a SKILL.md
Root-cause a failure and produce a minimal fix.
Structured debugging session — reproduce, isolate, diagnose, fix — driven by delegation. The main agent orchestrates while subagents do the investigation, repro, fix, and verification in parallel. Optional observability and source-control MCP enrichment.
Triggers: /debug {context and/or stacktrace and/or error information}, "debug this", "troubleshoot", "works locally but not in prod", flaky test, unexplained behavior
Cross-cutting workflows and general-purpose helpers.
Generic observability-to-ticket-to-branch workflow. Creates a ticket from an error in a connected observability MCP or from a free-form description, then opens a branch with git checkout -b <verb>/<TICKET-UPPER>, preserving uncommitted work. Works with any installed ticketing MCP (Shortcut, Linear, Jira, GitHub) and any observability MCP (Honeybadger, Sentry, Rollbar).
Triggers: /ticket-fix-flow <fault-synonym> <id>, /ticket-fix-flow <free-form description>
Converts webpages to markdown via markdown.new for efficient token usage.
Triggers: accessing, referencing, or analyzing content from public websites — not local-network or password-protected apps
Skills follow a specific structure with a required SKILL.md file containing YAML frontmatter and markdown instructions. See the create-skill skill for detailed guidance on creating effective skills.
skill-name/
├── SKILL.md (required)
│ ├── YAML frontmatter metadata (required)
│ │ ├── name: (required)
│ │ └── description: (required)
│ └── Markdown instructions (required)
└── Bundled Resources (optional)
├── scripts/ - Executable code
├── references/ - Documentation
└── assets/ - Files used in output
Skills are subject to routine security audits. While we strive to maintain a safe ecosystem, we encourage you to review skills before installation and use your own judgment. Report security issues to security.vercel.com.
See LICENSE for complete terms.